TobyOnFitnessTech
←All Episodes
AgentStack Daily: RSA Puts AI Agents on the Identity Roster with Agent ID — Episode 119 cover art
Episode 119·October 3, 2026·30:58

AgentStack Daily: RSA Puts AI Agents on the Identity Roster with Agent ID

RSA Puts AI Agents on the Identity Roster with Agent ID, Transformers 5.18 adds open-weight streaming speaker diarization, Firecrawl's MCP Server Turns Any... Show notes: https://tobyonfitnesstech.com/podcasts/episode-119/

🎧 Listen to Episode

AgentStack Daily EP119 — RSA Puts AI Agents on the Identity Roste, Transformers 5.18 adds open-weight strea, Firecrawl's MCP Server Turns Any LLM Int

Title: AgentStack Daily: RSA Puts AI Agents on the Identity Roster with Agent ID

Tagline: Today's stories: RSA Puts AI Agents on the Identity Roster with Agent ID, Transformers 5.18 adds open-weight streaming speaker diarization, Firecrawl's MCP Server Turns Any LLM Into a Web Scraper, and VDURA V12 Ships Multi-Tenant Storage for AI Factories. Concrete changes across the agent stack — what shipped, the mechanisms underneath, and what each one means for builders working with coding agents, models, and tooling.

Feed description: RSA Puts AI Agents on the Identity Roster with Agent ID, Transformers 5.18 adds open-weight streaming speaker diarization, Firecrawl's MCP Server Turns Any LLM Into a Web Scraper, and VDURA V12 Ships Multi-Tenant Storage for AI Factories. What shipped, how the mechanisms work, and what each change means for agent builders.

Story Slate

  1. RSA Puts AI Agents on the Identity Roster with Agent ID RSA launched Agent ID at The AI Conference in San Francisco, an identity security platform aimed at finance, government, healthcare, and critical infrastructure. It ships as three modules: Discover finds every agent and MCP server in the environment, Secure checks every tool call against policy in real time, and Govern maps evidence to ten regulatory frameworks. Discover and Secure go generally available November 16, 2026, with Govern following in the first half of 2027. Technical depth angle: Secure acts as an inline AI/MCP gateway that inspects every tool call at tool and argument depth. A risk engine scores each action on the user, the action type, and the target data sensitivity, so only high-risk calls route to a human via an out-of-band authenticated channel. Delegated permissions are inherited and never expanded, blocking privilege escalation through sub-agent handoffs. Actionability angle: What this means for builders: every agent needs a named human owner, a risk tier, and a kill switch before it touches a system of record. Why this matters now: a single badly worded prompt can take down an entire production instance, which is why policy checks need to live at the tool-call layer rather than in the prompt itself. Listener hook: One badly worded prompt took down an entire company's Salesforce — and most enterprises cannot name the agents already running inside their walls.

  2. Transformers 5.18 adds open-weight streaming speaker diarization Hugging Face shipped Transformers v5.18.0 as a stable release on September 30, 2026. The headline change is the addition of Nemotron 3 Diarization, an open-weight streaming model designed to determine who spoke when in real-world audio. Per the release notes, it supports both streaming and offline inference and is folded into the standard Transformers model registry. The release lands as a regular stable update on GitHub. Technical depth angle: Nemotron 3 Diarization is an open-weight model that takes in audio and assigns a speaker label to each segment. Streaming mode lets it produce those labels in real time as audio arrives, and offline mode runs the same weights against a pre-recorded file. Because it is added to the Transformers model registry, it loads through the same pipeline interface developers already use for other Hugging Face checkpoints. Actionability angle: Local self-hosters can now pull a speaker diarization model directly through Transformers instead of wiring up a separate diarization service, which makes it practical to attach who-spoke-when labels to transcripts and meeting notes without sending audio to a third-party API. Pairing it with a separate transcription model is still the typical setup for a full who-said-what record, since diarization does not produce text on its own. Listener hook: If you self-host audio pipelines, you can now get open-weight who-spoke-when labels without paying a third-party diarization service.

  3. Firecrawl's MCP Server Turns Any LLM Into a Web Scraper Firecrawl shipped an official Model Context Protocol server that gives Cursor, Claude, and other LLM clients the ability to scrape websites and search the web on demand. The open-source project has crossed 7,500 GitHub stars, and its latest release, v3.2.1, is now live. It lets an AI assistant pull clean text from any URL without you copy-pasting, which is a small change with big workflow implications for anyone doing competitive research, lead generation, or content monitoring. Technical depth angle: An MCP server is a plug-in that exposes tools to any LLM client speaking the Model Context Protocol. Firecrawl's server exposes two capabilities: one that fetches a URL and returns clean markdown, and one that runs a web search. Because MCP is an open standard, the same server plugs into Cursor, Claude Desktop, and other compatible clients with one install. Actionability angle: This means anyone using Cursor or Claude Desktop can now ask their assistant to fetch and summarize a webpage, pull structured data from a site, or run a web search mid-conversation without leaving the chat. Research and data-gathering steps that used to require manual browser work can now be folded into a single prompt. Listener hook: Your AI assistant can now read the web for you.

  4. VDURA V12 Ships Multi-Tenant Storage for AI Factories VDURA announced general availability of its Data Platform V12 on September 30, 2026, packaging multi-tenancy, API-first automation, and Context-Aware Tiering into a storage system now qualified on Supermicro hardware. The Pittsburgh and Abu Dhabi-based company, which serves AI factories and neocloud operators, claims more than double the performance per watt versus the previous generation. Technical depth angle: Context-Aware Tiering moves data across storage tiers based on access patterns, so frequently used datasets sit on fast media while older data shifts to denser, cheaper storage automatically, without manual policy work. Actionability angle: Neocloud and AI factory operators running mixed tenants on shared infrastructure now have a storage platform with built-in isolation and an API surface for automation. This matters because training and inference workloads benefit from tiering that adapts to how data is actually being used, and automating the split cuts both power and capacity costs. Listener hook: If you've ever wondered where the massive datasets that train and serve modern models actually live, this is one answer.

  5. Research digest: Self-Training AI Agents Can Quietly Drift Into Shared Blind Spots A new paper identifies a failure mode in AI systems that train themselves. When an agent generates its own practice questions and answers them in a loop, the question-generator and the answer-generator can gradually develop matching blind spots — agreeing on wrong answers that look right to themselves. Internal scores climb, but actual performance does not. The team calls this "co-cheating" and proposes a post-hoc audit against ground-truth evidence to catch it. The finding matters because more labs are using self-evolving training loops to bootstrap agent capabilities, and the failure is invisible without external verification. Technical depth angle: The paper studies self-evolving search agents that jointly train a question-generator and an answer-generator in a closed loop. The headline finding is that both components converge on shared errors — internal reward signals rise while external accuracy stalls. The proposed mitigation is auditing generated training data against original source documents after the fact, so shared blind spots surface before they corrupt the next training round. Plain English: the system grades itself, and once it agrees with itself on the wrong answer, the score stops meaning anything. Actionability angle: What this means for any self-training pipeline: an agent that grades its own work and trains on those grades needs an external reference signal, or internal scores can climb while real performance stays flat. Why this matters: the failure is invisible from inside the loop, so without an audit against source evidence, a self-evolving agent can quietly become better at agreeing with its own mistakes. Listener hook: If your AI agent is grading itself, it might be quietly training into a closed loop where it gets better at being wrong — together with itself.

  6. Google's Gemini 4 Argon hits 1M output tokens, gated to cyber defenders Google announced Gemini 4 Argon as a new frontier model with a 1 million token output limit, up from 64K. Pricing is set at $2 per million input tokens and $10 per million output tokens, with cached input priced at 95% off. Access is limited today to government testers and trusted cyber defenders through the Fairwind Program, with broader release pending safety work. Technical depth angle: The headline mechanic is a 1 million token output ceiling, up from 64K, which lets Argon sustain chain-of-thought reasoning across very long trajectories. Google pairs this with a phased rollout where cyber defenders and government users get access first via the Fairwind Program, before developers, enterprises, and consumers. Actionability angle: For most builders, this is a watch-and-wait moment: Argon isn't publicly available yet, only trusted cyber defenders through the Fairwind Program. Once Google opens access, the 1M-token output window and the long-horizon software engineering claims could matter to anyone building agentic workflows. Listener hook: Google's new Gemini 4 Argon already rewrote 32K lines of SIMD code into Rust on its own — and it has a million-token output window for thinking.

  7. Research digest: MemLife turns months of first-person video into searchable AI memory A research team built MemLife, a memory system that compresses hundreds of hours of first-person video into compact, searchable text summaries tied to the people, places, and objects the wearer encountered. When asked a question, a retrieval agent searches the memory timeline rather than reprocessing raw footage, keeping answers fast even as the archive grows. Without retraining, MemLife beat the strongest prior training-free baseline by 4.6 to 12 percent on four benchmarks that span months of video history. A companion method, MemOpt, applies reinforcement learning to make the memory summaries more faithful and easier to retrieve later. The work points toward AI assistants that genuinely remember a person's daily life instead of forgetting everything past the last few minutes. Technical depth angle: The system condenses wearable first-person video into short text summaries organized around the people, places, and objects the wearer saw, then a retrieval agent searches those summaries on a timeline to answer questions without reprocessing raw footage. A reinforcement-learning step then tunes how the summaries are written to make them more faithful and easier to find later. Plain English: it turns months of wearable camera footage into a searchable diary of who did what when, and learns to write that diary better over time. Actionability angle: This matters for builders working on AI companions, accessibility tools, or productivity apps on wearable cameras and smart glasses, since long-term recall is the missing layer between today's session-bound assistants and ones that genuinely know a user's history. The approach also signals where personal AI is heading: condensing continuous life data into a structured, searchable text layer instead of reprocessing raw streams on every question. The interesting downstream move would be extending the same memory-condensation idea to other personal data streams like audio logs or location trails. Listener hook: If you've ever wished an AI assistant could remember what you did last Tuesday instead of starting fresh each session, this is one of the closest working steps toward that yet.

  8. OpenAI Disrupts a Coordinated Effort to Extract Its Models' Reasoning OpenAI announced on September 30 that it disrupted a coordinated campaign to extract the reasoning behavior of its protected models through model distillation, a technique where one AI system is trained to imitate another's responses by sending many queries and learning from the answers. The company shut down the operation and is hardening defenses against similar adversarial efforts. Technical depth angle: Model distillation trains a student model to imitate a target model's outputs by sending many queries and learning from the answers. OpenAI says a coordinated version of that ran against its reasoning models to harvest their behavior, and the company is now hardening detection and response. Actionability angle: What this means: frontier reasoning models behind paid APIs are being actively defended as intellectual property, not just served. Why this matters: anyone assuming API outputs are unlimited training feedstock should expect probing patterns to be monitored, which makes the rules around legitimate fine-tuning worth watching closely. Listener hook: An organized group was systematically trying to extract how OpenAI's reasoning models think, and OpenAI says it shut them down.

  9. OpenAI partners with America's SBDC to bring hands-on AI help to small businesses OpenAI is teaming up with America's Small Business Development Center to expand hands-on AI training and local support for small businesses, alongside a new report on how small teams are actually putting AI to work. The program leans on SBDC's existing nationwide network of local advisors to meet business owners where they already show up for help. The aim is practical, in-person guidance rather than another generic webinar series. Technical depth angle: OpenAI is extending its reach through the SBDC's existing local advisor network rather than building a new training pipeline from scratch. The accompanying report is positioned as a snapshot of real small-team AI use, giving those sessions concrete examples to work from instead of abstract demos. Actionability angle: What this means: small business owners will likely start seeing hands-on AI sessions at their local SBDC office. For builders and toolmakers who sell to local businesses, the signal is that a more AI-literate customer base is about to walk through the door with practical questions about which tools to try first. Listener hook: If you run or advise a small business, free hands-on AI help may soon arrive at your local SBDC.

  10. Perplexity's Photon Cuts Search Latency 12x With a Rust Rewrite Perplexity has shipped Photon, an in-house retrieval and ranking engine written in Rust that now handles every production search request on its platform. The rewrite replaces a previously forked open-source engine, drops p99 latency from 800 ms to 65 ms, and powers a new Fast Search mode in the Perplexity Search API. For builders, that tail-latency win matters anywhere answers need to come back fast — chat, agents, autocomplete. Technical depth angle: The Rust rewrite consolidated retrieval and ranking into a single custom engine, replacing a forked open-source system. Owning both halves of the pipeline let Perplexity collapse two systems into one and trim p99 latency by roughly 12×. Actionability angle: What this means: builders using the Perplexity Search API can now route requests through the new Fast Search mode for speed-critical queries where tail latency matters. Why this matters: it shows the retrieval layer under the model — not the LLM itself — is often the real bottleneck for response time. Listener hook: If you've ever waited for a search box to think, Photon is what faster actually looks like.

  11. OpenAI introduces dots, proactive assistants that keep working while you step away OpenAI introduced dots on September 29, 2026, describing them as proactive assistants designed to keep working across complex projects and everyday tasks. The announcement frames dots as helpers that move work forward while keeping the user in control, signaling OpenAI's push toward assistants that continue past a single exchange. Technical depth angle: Dots are framed as proactive assistants that keep working across complex projects and everyday tasks rather than stopping after one turn, while still helping the user stay in control. Actionability angle: OpenAI is signaling a move toward assistants that keep moving work forward between prompts. For builders and users, the framing is less driving every step and more reviewing what got done while you were doing something else. The next thing to watch is whether dots actually deliver on the promise of keeping the user in control as work keeps moving. Listener hook: OpenAI just previewed a new kind of assistant that keeps going after your message ends.

  12. OpenAI Apologises to Australia, Pledges Stronger Cyber Safeguards OpenAI has apologised to Australia and committed to stronger safeguards following incidents involving Australian government websites. In a post published on September 28, 2026, the company said it is rolling out additional support to help strengthen Australia's cyber defences. The announcement, titled "How we will do better for Australia," is positioned as a reset in how OpenAI engages with Australian public-sector customers, pairing an apology with forward-looking safeguards. What those safeguards look like in practice, and how the specific incidents will be addressed, will shape how seriously Australian agencies treat the pledge. Technical depth angle: This is a policy and relationship announcement rather than a technical release. The post frames its commitments around safeguards and cyber-defence support, with no new model versions, APIs, or product features named in the announcement itself. Actionability angle: For builders, the practical effect today is limited — there is no new API surface, model, or integration to adopt. Anyone serving Australian public-sector clients should watch for follow-up posts that translate the "stronger safeguards" pledge into concrete controls, because right now the announcement reads as a commitment rather than a deployed change. Listener hook: OpenAI is on the back foot in Australia after incidents tied to government sites, and is publicly promising to do better.

  13. OpenAI publishes early guidelines for safety cases in frontier training OpenAI published early-stage guidelines for building safety cases around frontier AI training runs on September 28, 2026. The guidelines organize the safety argument into three pillars: the technical safeguards in place during training, the operational practices supporting them day-to-day, and the procedures used to investigate a misalignment incident if one surfaces. It is framed as an early working document rather than a finished standard. Technical depth angle: A safety case is the structured argument that a training run meets a defined safety bar, and OpenAI's draft splits that argument into three pillars: technical controls during training, day-to-day operational practices supporting them, and an incident-investigation playbook for misalignment when a model behaves in unintended ways. Actionability angle: For builders outside a frontier lab, the document is mostly a read-not-implement signal: it shows what safety documentation requests may start to look like if you ship safety-critical features for a frontier partner. Whether other labs publish similar frameworks is the next thing to watch. Listener hook: OpenAI just sketched, in plain language, what a safety case for a frontier training run actually looks like, and it is worth a look if you build anything that touches a frontier model.

  14. Microsoft's Quine takes aim at biology's data sprawl Microsoft Research has unveiled Quine, an early-stage AI effort designed to build a multimodal world model of biology. The system aims to connect evidence across biological scales and data types so researchers can computationally search hypothesis spaces far larger than intuition allows, then prioritize the most promising candidates for the lab. Experimental results flow back into the model to sharpen future directions. Technical depth angle: Quine is framed as a multimodal world model that links biological insights across scales and modalities, so researchers can computationally narrow huge hypothesis spaces and feed lab results back into the next round of inquiry. Actionability angle: For biologists and research-tool builders, the practical signal is that Microsoft is investing in systems that unify diverse biological data rather than treating each data layer as its own silo. For now, Quine is a research effort rather than a shipping product, so the immediate opportunity is to watch for partner calls or data-format standards worth aligning with. The choice of which biological modality Quine tackles first will shape what tooling becomes worth building around. Listener hook: AI is finally getting a serious attempt at modeling life itself — and Microsoft wants to start by mapping biology, not just text.

Editorial Mix Check

  • flagship_products: 5
  • builder_projects: 6
  • local_ai: 2
  • hardware_compute: 2
  • policy_regulation: 4
  • research: 2

Model Discovery Check

  • OpenAI: GPT-6.1 Sol Pro (openai) — Newly listed this cycle (verified October 01, 2026). Primary source: https://openrouter.ai/models/openai/gpt-6.1-sol-pro. Availability: API via OpenRouter. params_active: n/a; params_total: n/a; context: 1050000 tokens; modality: see primary source. Capabilities: context length 1050000; GPT-6.1 Sol Pro is the same underlying model as GPT-6.1 Sol, served with reasoning.mode set to pro for higher-quality responses on complex tasks. Cost note: pro mode sp. Try now / integration angle: Route a coding-agent session through https://openrouter.ai/models/openai/gpt-6.1-sol-pro and compare it with the current default. Decision: Selected — new major-provider model not featured on a recent broadcast.

  • OpenAI: GPT-6.1 Sol (openai) — Newly listed this cycle (verified October 01, 2026). Primary source: https://openrouter.ai/models/openai/gpt-6.1-sol. Availability: API via OpenRouter. params_active: n/a; params_total: n/a; context: 1050000 tokens; modality: see primary source. Capabilities: context length 1050000; GPT-6.1 Sol is an upgrade to GPT-6 Sol from OpenAI, positioned below the flagship GPT-6 Astra in the GPT-6 series. It is suited for agentic coding, computer use, document-heavy professional.... Try now / integration angle: Route a coding-agent session through https://openrouter.ai/models/openai/gpt-6.1-sol and compare it with the current default. Decision: Selected — new major-provider model not featured on a recent broadcast.

Local LLM Spotlight

  • Edge0/Audio8-ASR-Infinite — https://huggingface.co/Edge0/Audio8-ASR-Infinite — Trending open model on Hugging Face; task automatic-speech-recognition; 1986 likes and 27228 downloads. Tags: transformers, safetensors, audio8_asr_infinite, text-generation, streaming, realtime, speech-recognition, audio, automatic-speech-recognition, custom_code. Try now: Read the linked model card before downloading, and choose a runtime or device only after it confirms the license, weight format, context window, benchmarks, and hardware requirements.

GitHub Project Radar

  • HKUDS/nanobot — https://github.com/HKUDS/nanobot — Ultra-lightweight, open-source, self-hosted personal AI agent framework in Python with WebUI, tools, memory, MCP, multi-agent workflows, automation, and chat apps stars: 48,712; stars_delta_30d: +1,114 (+2.3%) since 2026-09-01; latest_release: v0.3.5 (2026-09-15). Why this is on the radar now: v0.3.5 shipped on 2026-09-15 and the repository was updated on 2026-10-01. Stack improvement angle: Adds a tool surface MCP-compatible agents (OpenClaw, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.

  • DeusData/codebase-memory-mcp — https://github.com/DeusData/codebase-memory-mcp — High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary stars: 45,593; stars_delta_30d: +3,997 (+9.6%) since 2026-09-01; latest_release: v0.11.0 (2026-09-15). Why this is on the radar now: v0.11.0 shipped on 2026-09-15 and the repository was updated on 2026-09-30. Stack improvement angle: Adds a tool surface MCP-compatible agents (OpenClaw, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.

  • ahujasid/mcp-for-blender — https://github.com/ahujasid/mcp-for-blender — Community plugin to control Blender 3D with any LLM of your choice stars: 29,789; stars_delta_30d: n/a — first tracked appearance; latest_release: none published on GitHub as of 2026-10-01. Why this is on the radar now: The repository was updated on 2026-09-30 and enters the radar with 29,789 stars. Stack improvement angle: Adds a tool surface MCP-compatible agents (OpenClaw, Codex, Claude Code, Hermes) can call directly. Try now: Clone the repo and wire it into a test agent session to evaluate the tool surface.

Extra Research Candidates

Show Notes

Episode 119 — October 01, 2026

[00:00] Episode hook

RSA Puts AI Agents on the Identity Roster with Agent ID headlines a dense cycle. Transformers 5.18 adds open-weight streaming speaker diarization, Firecrawl's MCP Server Turns Any LLM Into a Web Scraper, VDURA V12 Ships Multi-Tenant Storage for AI Factories round out the front of the episode, with deeper cuts across models, tooling, and infrastructure behind them. Each story gets the same treatment — what shipped, the mechanism underneath, and what it changes for working builders.

[02:00] RSA Puts AI Agents on the Identity Roster with Agent ID

RSA launched Agent ID at The AI Conference in San Francisco. President and Chief Product and Strategy Officer Jim Taylor framed the problem plainly: AI agents are not service accounts. They are dynamic, accumulate permissions, and rarely have an owner.

The scale is already past guessing. Gartner expects a typical Global Fortune 500 enterprise to run roughly 150,000 AI agents by 2028, up from fewer than 15 in 2025, while only 13% of organizations believe they have the right agent governance in place. RSA's audit at a medium-sized global bank — whose policy prohibited agents — found more than 4,000.

Taylor's failure story needed no attacker. A customer service employee at an unnamed company asked an agent to "go to Salesforce and get all the data" for health charts. The agent downloaded the database. Salesforce flagged the traffic as a denial-of-service attack and shut the instance down. One prompt, one operator, one company outage.

Agent ID ships as three modules. Discover scans endpoints, devices, networks, and apps through CrowdStrike and Zscaler connectors, then registers every agent and MCP server as a first-class identity with a named owner, risk tier, and lifecycle state, linked to providers like Microsoft Entra ID, Okta, and AWS IAM. Secure sits inline as an AI/MCP gateway, evaluating every tool call at tool and argument depth — allowing, denying, or escalating to the owner through an out-of-band, phishing-resistant channel. Govern logs every action and maps evidence to ten regulatory frameworks, streaming it to the customer's SIEM.

The delegation model closes a privilege-escalation path: an agent can only enable another agent with the entitlements it was granted, never expanding inherited permissions. Approvals run through a risk engine scoring user, action, and target endpoint — a refund under $500 might process automatically while a larger one needs a second approver.

Discover and Secure ship generally available November 16, 2026. Govern follows in the first half of 2027.

[02:51] Transformers 5.18 adds open-weight streaming speaker diarization

Hugging Face shipped Transformers v5.18.0 as a stable release on September 30, 2026. The release notes announce one headline addition: Nemotron 3 Diarization, an open-weight streaming model built to determine who spoke when in real-world audio.

Per the published notes, the model supports both streaming and offline inference, which means it can assign speaker labels in real time as audio arrives or run against a pre-recorded file. The release notes excerpt indicates it handles up to eight, though the published text is truncated at that point. Because the weights are open, local self-hosters can run the model on their own hardware rather than calling a hosted diarization service.

Adding the model to the Transformers registry means it loads through the same pipeline interface developers already use for other Hugging Face checkpoints. There is no new API surface to learn; it is a new entry in the existing model zoo.

For builders working on local audio pipelines, the practical unlock is straightforward: speaker labels attached to recorded audio without sending the audio to a third-party service. Diarization is not a speech-to-text model, so pairing it with a separate transcription model is still the typical setup for a complete who-said-what record.

That is the scope of v5.18.0 as published in the source notes: one new open-weight model for streaming and offline speaker diarization, available through the standard Transformers pipeline interface.

[04:17] Firecrawl's MCP Server Turns Any LLM Into a Web Scraper

Firecrawl's official MCP server crossed 7,500 GitHub stars with its latest release, v3.2.1, giving Cursor, Claude, and any other Model Context Protocol-compatible LLM client the ability to scrape websites and search the web on demand.

An MCP server is a small plug-in that exposes tools to an AI assistant using the Model Context Protocol, the open standard that lets LLMs reach outside their chat window into real data and services. Firecrawl's server exposes two capabilities: one that pulls a URL and returns clean markdown, and one that runs a web search. Because MCP is a standard, the same server plugs into Cursor, Claude Desktop, and other compatible clients with a single install — no custom integration per app.

For builders, the practical shift is that research and data-gathering steps that used to mean opening a browser can now happen inside the conversation. You can ask your coding assistant to fetch documentation from a vendor site and summarize it, or ask Claude to pull pricing tables from competitor pages and turn them into structured notes. Anything you'd otherwise copy-paste between a browser tab and your chat becomes a single prompt.

The project is open source and sits at v3.2.1. For nearly any AI workflow that needs fresh or external information, Firecrawl's MCP server removes the browser hop.

[05:39] VDURA V12 Ships Multi-Tenant Storage for AI Factories

VDURA, a Pittsburgh and Abu Dhabi-based data storage company that serves neoclouds and AI factories, announced general availability of Data Platform V12 on September 30. The release bundles four pieces aimed at AI infrastructure operators: multi-tenancy for shared infrastructure, an API-first automation surface so teams can script storage operations, Context-Aware Tiering that moves data between storage tiers based on access patterns, and a claim of more than double the performance per watt compared with the prior generation. V12 is also now qualified on Supermicro building blocks, giving buyers a pre-validated hardware path rather than a custom integration. Context-Aware Tiering is the most concrete piece of new behavior: frequently used datasets stay on fast drives, while older data migrates to denser, cheaper media automatically, without manual policy work. For AI factories running many tenants on shared hardware, the multi-tenancy plus API combination means storage can be provisioned and rebalanced programmatically rather than through tickets. The watt-efficiency claim matters because storage at this scale draws real energy, and a doubling of performance per watt is the kind of number an infrastructure team can budget around.

[06:48] Research digest: Self-Training AI Agents Can Quietly Drift Into Shared Blind Spots

When AI agents train themselves, they can quietly drift into shared blind spots. A new paper studies what's called self-evolving search agents — systems that build their own practice questions and then answer them, in a loop. One component generates the questions. Another tries to answer them. They score each other, refine, repeat.

The team flags a failure mode they name co-cheating: the question-generator and the answer-generator start agreeing on wrong answers that look plausible to both. Internal reward goes up. Actual accuracy does not. And it gets worse the longer the loop runs.

The proposed fix is a post-hoc audit — checking the generated training data against the original source documents the agent was supposed to be learning from, after the fact. The audit surfaces where both halves of the loop quietly locked onto the same error.

The practical takeaway: if you are building any system that grades its own output and trains on that grade, you need an outside reference signal. Otherwise the score can rise while the model just gets better at agreeing with itself on the wrong thing.

[07:57] Google's Gemini 4 Argon hits 1M output tokens, gated to cyber defenders

Google today unveiled Gemini 4 Argon, a frontier model with a 1 million token output limit, up from 64K. The new ceiling lets the model sustain chain-of-thought reasoning across much longer trajectories, which Google says unlocks deeper multi-step problem solving in coding, financial research, legal drafting, and autonomous cybersecurity work.

Pricing is set: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off.

Right now, access is restricted. Argon is rolling out through Google's Fairwind Program to government users and trusted cyber defenders. Wider availability is pending further safety testing under Google's Frontier Safety Framework. Google is also engaged in the U.S. government's voluntary process for pre-release model access.

Inside Google, the model is already in production. Engineers report using Argon for quantum algorithm optimization — in one case beating a published baseline by 40% in minutes. Agent fleets analyzed data center telemetry and freed over 300 TiB of memory. The most striking example: agents migrating C/C++ codebases to Rust, including re2, libgav1, and Fuchsia's Zircon kernel at 800K+ lines. On libgav1, agents replaced 32K lines of SIMD code with safe Rust and produced a memory-safe decoder that runs 2.7x faster than the prior Rust port.

On benchmarks, Argon hits 77.9% on DeepSWE v1.1, leads the Vals Index across finance, legal, and tax work, ranks #1 on Zapier's AutomationBench at 51.3%, scores 91.7% on LVBench for long video understanding, and ties for first on CWE-bench v1 at 68%. For cyber defense, trusted testers get the model without cyber guardrails. Wiz is already using Argon through its Scan for Good initiative and uncovered a critical exposure that prior frontier models missed.

Watch next — when Argon actually opens to developers, and what the Fairwind cyber trials reveal.

[09:48] Research digest: MemLife turns months of first-person video into searchable AI memory

Imagine wearing a camera that captures your entire day, every day, for months. An AI assistant could later answer "what did I cook for dinner last Tuesday?" or "did the plumber say anything about the water heater?" A research team has taken a real step toward that future with MemLife, a memory system that condenses hundreds of hours of first-person video into compact text summaries keyed to the people, places, and objects the wearer encountered. When a query arrives, a retrieval agent searches the memory timeline instead of reprocessing raw footage, keeping answers fast as the archive grows. Without retraining, MemLife outperformed the strongest prior training-free approach by 4.6 to 12 percent across four benchmarks that span months of video history. A second piece, MemOpt, uses reinforcement learning to tune the memory writer so its summaries stay faithful and easy to find later. Together they point toward AI companions that genuinely remember your life, not just the last few minutes of conversation.

[10:49] OpenAI Disrupts a Coordinated Effort to Extract Its Models' Reasoning

OpenAI announced on September 30 that it disrupted a coordinated campaign aimed at extracting the reasoning behavior of its protected models. The effort involved model distillation, a technique where one AI system is trained to imitate another's behavior by sending many queries and learning from the answers.

The word OpenAI used was "coordinated," not "individual," which frames this as an organized operation rather than a lone experimenter probing the API. That distinction matters because distillation at scale requires automation, and automation leaves traces that defenders can detect.

OpenAI says it shut the campaign down and is hardening defenses against adversarial distillation, the practice of training a competing model by systematically extracting behavior from a target. The company is treating its reasoning models as intellectual property worth active defense.

For builders, the practical takeaway is that frontier reasoning models are being monitored in real time. Anyone planning to fine-tune a model on the output of a paid API should expect that usage pattern to be visible and enforceable.

One thing to watch: whether OpenAI publishes more on how coordinated probing campaigns get detected, because the defensive playbook for adversarial distillation matters to anyone running their own hosted model.

[12:03] OpenAI partners with America's SBDC to bring hands-on AI help to small businesses

OpenAI announced a partnership with America's Small Business Development Center to bring hands-on AI training and local support to small businesses, paired with a new report on how small teams are putting AI to work. The announcement landed on September 30, 2026, and it leans on the SBDC's existing nationwide network of local advisors, the same people small business owners already visit for help with plans, loans, and growth questions.

The approach is straightforward: rather than build a brand new training pipeline, OpenAI is plugging into a network that already meets business owners in their own communities. That means hands-on training and local support delivered in person by advisors who know the local economy, not a generic webinar series. The accompanying report is meant to give those sessions real grounding by documenting how small teams are actually using AI today, so advisors can show what is already working for teams of a handful of people rather than enterprise-scale deployments.

For small business owners, the practical takeaway is that their local SBDC is likely to start offering hands-on sessions on putting AI to work in day-to-day operations. For builders and toolmakers who sell to local businesses, the signal is that a more AI-literate customer base is about to walk through the door. One thing to watch next: which SBDC regions roll out the program first and what concrete examples from the report get used in those first sessions.

[13:33] Perplexity's Photon Cuts Search Latency 12x With a Rust Rewrite

Perplexity just shipped Photon, a retrieval system it wrote from scratch in Rust, and it now handles every search request going through the company's AI search stack. That includes the consumer product and the developer-facing Search API.

The headline number is latency. Photon reportedly cuts p99 latency — the response time for the slowest 1% of queries — from 800 milliseconds down to 65 milliseconds. That's roughly a 12× improvement on the tail, which is where users actually feel lag.

Photon replaces an open-source engine Perplexity had previously forked and customized. Rather than keep patching someone else's code, the team rewrote the retrieval and ranking pipeline in Rust, a systems language known for tight memory control and fast threading. By owning the whole stack, Perplexity could collapse retrieval and ranking into a single engine instead of stitching two systems together.

For developers, the immediate change is a new Fast Search mode in the Perplexity Search API. If you're building anything that needs answers back fast — live chat, agent loops, autocomplete — this is the mode aimed at you.

How it shipped also says something about the layer under the model. Most of the public attention goes to which LLM a company uses. Photon is a reminder that the search layer underneath can be just as much of a bottleneck, and rewriting it in a systems language is one of the few ways to win big on latency without throwing more hardware at the problem.

One thing worth watching next: whether Perplexity opens any of Photon up. A Rust retrieval engine with that kind of latency profile would interest a lot of teams building their own search-heavy products.

[15:18] OpenAI introduces dots, proactive assistants that keep working while you step away

OpenAI introduced dots on September 29, 2026, describing them as proactive assistants that can keep working across complex projects and everyday tasks. The framing in OpenAI's own announcement centers on the idea that dots help you stay in control while the work moves forward, suggesting an assistant that continues across work rather than stopping after each exchange. OpenAI positions dots as useful for both multi-step projects and ordinary everyday tasks, without specifying pricing, platform availability, or the underlying technical mechanism in the announcement. That gap matters, because this is OpenAI's own framing of what dots do, not a feature checklist or spec sheet. Anyone waiting to see how dots handle long-running tasks in practice will want hands-on details once people start using them, since staying in control while work moves forward is a promise rather than a confirmed workflow.

[16:10] OpenAI Apologises to Australia, Pledges Stronger Cyber Safeguards

OpenAI has apologised to Australia and committed to stronger safeguards after incidents involving Australian government websites, in a post published on September 28, 2026. The announcement, titled "How we will do better for Australia," runs through OpenAI's official news channel and frames the company as a partner willing to harden its posture for Australian public-sector customers. The core of the post is a two-part move: an apology, and a forward-looking pledge that includes stronger safeguards and additional support aimed at strengthening Australia's cyber defences. That makes the announcement read as a relationship reset, not a product release. There is no new model, no new API, and no integration story to chase — the work is about how OpenAI operates inside an Australian government context, and about rebuilding confidence after the incidents the company is now responding to. For Australian government and public-sector teams already using OpenAI tools, the immediate question is whether the promised safeguards will arrive as new technical controls, new account-level options, or new contractual terms. For everyone else, the episode is a useful reminder that frontier labs operate inside national regulatory and political contexts, and that a country's relationship with a provider can shift on incidents the broader market barely registers. The thing to watch next is the follow-up — whether OpenAI publishes a more concrete technical or policy document that turns the "stronger safeguards" line into something Australian agencies can point to, or whether the pledge stays at the level of a public commitment.

[17:44] OpenAI publishes early guidelines for safety cases in frontier training

OpenAI released early guidelines for safety cases in frontier AI training on September 28. The document sketches what a structured safety argument around a major training run could look like, framed as a working draft rather than a finished standard.

The framework rests on three pillars. The first is the technical safeguards in place during training, which covers the controls that gate what a model can and cannot do while it is being built. The second is the operational practices supporting those safeguards day-to-day, the human and procedural side of keeping the controls working. The third is an incident playbook for investigating misalignment when a model behaves in ways its developers did not intend.

For most builders, the direct impact is limited. The document is a read signal: it shows what frontier labs are starting to expect by way of structured safety arguments, and what a partner safety review may start asking about on projects that touch frontier models.

[18:44] Microsoft's Quine takes aim at biology's data sprawl

Microsoft Research has introduced Quine, an early-stage research effort aimed at one of the messier targets in science: biology. The premise is that life doesn't operate in clean silos — a cell, a tissue, and a clinical outcome live in different data formats and different scales — so an AI built to model it shouldn't either.

Quine is described as a multimodal world model of biology. In plain English, that means a system designed to absorb and connect many kinds of biological evidence at once, rather than handling, say, genomics and imaging in separate pipelines. The goal, Microsoft says, is to let scientists computationally search a hypothesis space far larger than intuition allows, then prioritize the most promising candidates before committing lab time.

A key piece of the design loop is feedback. Experimental results don't just sit at the end — they get folded back in to sharpen future research directions. That iterative pattern is what turns a model from a static encyclopedia into something closer to a research partner.

Microsoft is positioning Quine as an early-stage effort, not a finished product. The post frames it as a foundation for scientists to explore biology computationally at scales no human could hold in their head, with the lab acting as the ground truth. The interesting thing to watch next is which biological modalities and which partner labs Microsoft chooses to seed the system with — that will determine whether Quine becomes a general research surface or a tool aimed at one corner of biology.

Chapters

  • 00:00 — Intro: RSA Puts AI Agents on the Identity Roster with Agent ID / Transformers 5.18 adds open-weight streaming speaker diarization / Firecrawl's MCP Server Turns Any LLM Into a Web Scraper
  • 02:00 — RSA Puts AI Agents on the Identity Roster with Agent ID
  • 02:51 — Transformers 5.18 adds open-weight streaming speaker diarization
  • 04:17 — Firecrawl's MCP Server Turns Any LLM Into a Web Scraper
  • 05:39 — VDURA V12 Ships Multi-Tenant Storage for AI Factories
  • 06:48 — Research digest: Self-Training AI Agents Can Quietly Drift Into Shared Blind Spots
  • 07:57 — Google's Gemini 4 Argon hits 1M output tokens, gated to cyber defenders
  • 09:48 — Research digest: MemLife turns months of first-person video into searchable AI memory
  • 10:49 — OpenAI Disrupts a Coordinated Effort to Extract Its Models' Reasoning
  • 12:03 — OpenAI partners with America's SBDC to bring hands-on AI help to small businesses
  • 13:33 — Perplexity's Photon Cuts Search Latency 12x With a Rust Rewrite
  • 15:18 — OpenAI introduces dots, proactive assistants that keep working while you step away
  • 16:10 — OpenAI Apologises to Australia, Pledges Stronger Cyber Safeguards
  • 17:44 — OpenAI publishes early guidelines for safety cases in frontier training
  • 18:44 — Microsoft's Quine takes aim at biology's data sprawl

Primary Links

Release Coverage Check

  • Hermes Agent — Latest stable verified: v2026.9.24, published 2026-09-24T10:09:38Z. Recent episode version tags detected: v2026.9.14, v2026.9.21, v2026.9.24, v2026.9.7. No new stable release this cycle.
  • OpenAI Codex — Latest stable verified: rust-v0.159.3, published 2026-09-30T22:57:34Z. Recent episode version tags detected: rust-v0.155.1, rust-v0.156.1, rust-v0.157.0, rust-v0.159.0. No new stable release this cycle.
  • Claude Code CLI — Latest stable verified: 2.1.285, published 2026-09-29T17:32:09.173Z. Recent episode version tags detected: 2.1.274, 2.1.277, latest, stable. No new stable release this cycle.
  • Antigravity CLI — Continuous delivery model; no discrete release tags verified this cycle (latest build as of 2026-10-01). Recent episode version tags detected: @google/antigravity, v2.0.

Harness Version Reference

  • Hermes Agent — v2026.9.24
  • OpenAI Codex — rust-v0.159.3
  • Claude Code CLI — 2.1.285
  • Antigravity CLI — Continuous delivery (no tagged release verified this cycle)

🎙 Never miss an episode — subscribe now

SpotifyAppleiHeartYouTube MusicAmazonRSS
🎙 Subscribe to AgentStack Daily